libdwfl: Rewrite reading of ar_size in elf_begin_rand
Commit Message
With GCC 12.1.1, glibc 2.3a, -fsanitize=undefined and
-D_FORTIFY_SOURCE=3 we get the following error message:
In file included from /usr/include/ar.h:22,
from ../libelf/libelfP.h:33,
from core-file.c:31:
In function ‘pread’,
inlined from ‘pread_retry’ at ../lib/system.h:188:21,
inlined from ‘elf_begin_rand’ at core-file.c:86:16,
inlined from ‘core_file_read_eagerly’ at core-file.c:205:15:
/usr/include/bits/unistd.h:74:10: error: ‘__pread_alias’ writing 58 or more bytes into a region of size 10 overflows the destination [-Werror=stringop-overflow=]
74 | return __glibc_fortify (pread, __nbytes, sizeof (char),
| ^~~~~~~~~~~~~~~
/usr/include/ar.h: In function ‘core_file_read_eagerly’:
/usr/include/ar.h:41:10: note: destination object ‘ar_size’ of size 10
41 | char ar_size[10]; /* File size, in ASCII decimal. */
| ^~~~~~~
/usr/include/bits/unistd.h:50:16: note: in a call to function ‘__pread_alias’ declared with attribute ‘access (write_only, 2, 3)’
50 | extern ssize_t __REDIRECT (__pread_alias,
| ^~~~~~~~~~
cc1: all warnings being treated as errors
The warning disappears when dropping either -fsanitize=undefined
or when using -D_FORTIFY_SOURCE=2. It looks like a false positive.
But I haven't figured out how/why it happens.
The code is a little tricky to proof correct though. The ar_size
field is a not-zero terminated string ASCII decimal, right-paddedr
with spaces. Which is then converted with strtoll. Relying on the
fact that the struct ar_hdr is zero initialized, so there will be
a zero byte after the ar_size field.
Rewrite the code to just use a zero byte terminated char array.
Which is much easier to reason about. As a bonus the error disappears.
Signed-off-by: Mark Wielaard <mark@klomp.org>
---
libdwfl/ChangeLog | 5 +++++
libdwfl/core-file.c | 26 ++++++++++++++++----------
2 files changed, 21 insertions(+), 10 deletions(-)
Comments
On Thu, 2022-07-28 at 15:48 +0200, Mark Wielaard wrote:
> With GCC 12.1.1, glibc 2.3a, -fsanitize=undefined and
> -D_FORTIFY_SOURCE=3 we get the following error message:
Sorry for the typo, it is glibc 2.35. Basically an up to date Fedora 36
system (replicated on x86_64, ppc64le and s390x).
On 2022-07-28 09:48, Mark Wielaard wrote:
> With GCC 12.1.1, glibc 2.3a, -fsanitize=undefined and
> -D_FORTIFY_SOURCE=3 we get the following error message:
>
> In file included from /usr/include/ar.h:22,
> from ../libelf/libelfP.h:33,
> from core-file.c:31:
> In function ‘pread’,
> inlined from ‘pread_retry’ at ../lib/system.h:188:21,
> inlined from ‘elf_begin_rand’ at core-file.c:86:16,
> inlined from ‘core_file_read_eagerly’ at core-file.c:205:15:
> /usr/include/bits/unistd.h:74:10: error: ‘__pread_alias’ writing 58 or more bytes into a region of size 10 overflows the destination [-Werror=stringop-overflow=]
> 74 | return __glibc_fortify (pread, __nbytes, sizeof (char),
> | ^~~~~~~~~~~~~~~
> /usr/include/ar.h: In function ‘core_file_read_eagerly’:
> /usr/include/ar.h:41:10: note: destination object ‘ar_size’ of size 10
> 41 | char ar_size[10]; /* File size, in ASCII decimal. */
> | ^~~~~~~
> /usr/include/bits/unistd.h:50:16: note: in a call to function ‘__pread_alias’ declared with attribute ‘access (write_only, 2, 3)’
> 50 | extern ssize_t __REDIRECT (__pread_alias,
> | ^~~~~~~~~~
> cc1: all warnings being treated as errors
>
> The warning disappears when dropping either -fsanitize=undefined
> or when using -D_FORTIFY_SOURCE=2. It looks like a false positive.
> But I haven't figured out how/why it happens.
Interesting, I'll take a closer look at this from the gcc context. I
obviously don't have any strong opinions about the elfutils patch :)
Thanks,
Sid
> The code is a little tricky to proof correct though. The ar_size
> field is a not-zero terminated string ASCII decimal, right-paddedr
> with spaces. Which is then converted with strtoll. Relying on the
> fact that the struct ar_hdr is zero initialized, so there will be
> a zero byte after the ar_size field.
>
> Rewrite the code to just use a zero byte terminated char array.
> Which is much easier to reason about. As a bonus the error disappears.
>
> Signed-off-by: Mark Wielaard <mark@klomp.org>
> ---
> libdwfl/ChangeLog | 5 +++++
> libdwfl/core-file.c | 26 ++++++++++++++++----------
> 2 files changed, 21 insertions(+), 10 deletions(-)
>
> diff --git a/libdwfl/ChangeLog b/libdwfl/ChangeLog
> index 75c53948..acdaa013 100644
> --- a/libdwfl/ChangeLog
> +++ b/libdwfl/ChangeLog
> @@ -1,3 +1,8 @@
> +2022-07-28 Mark Wielaard <mark@klomp.org>
> +
> + * core-file.c (elf_begin_rand): Replace struct ar_hdr h with
> + a char ar_size[AR_SIZE_CHARS + 1] array to read size.
> +
> 2022-07-18 Shahab Vahedi <shahab@synopsys.com>
>
> * debuginfod-client.c (dwfl_get_debuginfod_client stub):
> diff --git a/libdwfl/core-file.c b/libdwfl/core-file.c
> index cefc3db0..4418ef33 100644
> --- a/libdwfl/core-file.c
> +++ b/libdwfl/core-file.c
> @@ -75,26 +75,32 @@ elf_begin_rand (Elf *parent, off_t offset, off_t size, off_t *next)
> from the archive header to override SIZE. */
> if (parent->kind == ELF_K_AR)
> {
> - struct ar_hdr h = { .ar_size = "" };
> -
> - if (unlikely (parent->maximum_size - offset < sizeof h))
> + /* File size, in ASCII decimal, right-padded with ASCII spaces.
> + Max 10 characters. Not zero terminated. So make this ar_size
> + array one larger and explicitly zero terminate it. As needed
> + for strtoll. */
> + #define AR_SIZE_CHARS 10
> + char ar_size[AR_SIZE_CHARS + 1];
> + ar_size[AR_SIZE_CHARS] = '\0';
> +
> + if (unlikely (parent->maximum_size - offset < sizeof (struct ar_hdr)))
> return fail (ELF_E_RANGE);
>
> if (parent->map_address != NULL)
> - memcpy (h.ar_size, parent->map_address + parent->start_offset + offset,
> - sizeof h.ar_size);
> + memcpy (ar_size, parent->map_address + parent->start_offset + offset,
> + AR_SIZE_CHARS);
> else if (unlikely (pread_retry (parent->fildes,
> - h.ar_size, sizeof (h.ar_size),
> + ar_size, AR_SIZE_CHARS,
> parent->start_offset + offset
> + offsetof (struct ar_hdr, ar_size))
> - != sizeof (h.ar_size)))
> + != AR_SIZE_CHARS))
> return fail (ELF_E_READ_ERROR);
>
> - offset += sizeof h;
> + offset += sizeof (struct ar_hdr);
>
> char *endp;
> - size = strtoll (h.ar_size, &endp, 10);
> - if (unlikely (endp == h.ar_size)
> + size = strtoll (ar_size, &endp, 10);
> + if (unlikely (endp == ar_size)
> || unlikely ((off_t) parent->maximum_size - offset < size))
> return fail (ELF_E_INVALID_ARCHIVE);
> }
On 2022-07-28 13:27, Siddhesh Poyarekar wrote:
> Interesting, I'll take a closer look at this from the gcc context. I
> obviously don't have any strong opinions about the elfutils patch :)
I reduced this to the below program and I see it warns with
`-D_FORTIFY_SOURCE=3 -O2 -fsanitize=undefined` as well as
`-D_FORTIFY_SOURCE=2 -O2 -fsanitize=undefined`. This is definitely a
false positive on unreachable code; __pread_alias will never be called
when nbytes is greater than object size.
So I can confirm that this is harmless.
Sid
~~~
typedef long ssize_t;
typedef unsigned long size_t;
typedef long off_t;
struct ar_hdr
{
char ar_mode;
char ar_size[10];
};
extern ssize_t __pread_chk (int fd, void *buf, size_t nbytes, off_t offset,
size_t bufsize)
__attribute__((__access__ (__write_only__, 2, 3)));
extern ssize_t __pread_alias (int fd, void *buf, size_t nbytes, off_t
offset)
__attribute__((__access__ (__write_only__, 2, 3)));
extern ssize_t __pread_chk_warn (int fd, void *buf, size_t nbytes,
off_t offset, size_t bufsize);
extern __inline __attribute__((__always_inline__))
__attribute__((__gnu_inline__)) ssize_t
pread (int fd, void *buf, size_t nbytes, off_t offset)
{
size_t osz = __builtin_dynamic_object_size (buf, 0);
if (__builtin_constant_p (osz) && osz == (size_t) -1)
return __pread_alias (fd, buf, nbytes, offset);
return (((__typeof (nbytes)) 0 < (__typeof (nbytes)) - 1
|| (__builtin_constant_p (nbytes) && (nbytes) > 0))
&& __builtin_constant_p (nbytes <= osz / 1)
&& nbytes <= osz / 1)
? __pread_alias (fd, buf, nbytes, offset)
: __pread_chk (fd, buf, nbytes, offset, osz);
}
ssize_t
pread_retry (int fd, off_t start_offset, off_t offset)
{
ssize_t recvd = 0;
struct ar_hdr h = {.ar_size = {0} };
void *buf = h.ar_size;
size_t len = sizeof (h.ar_size);
off_t off =
start_offset + offset + __builtin_offsetof (struct ar_hdr, ar_size);
do
{
long int res;
do
{
res = pread (fd, ((char *) buf) + recvd, len - recvd, off +
recvd);
}
while (res == -1L);
recvd += res;
}
while ((size_t) recvd < len);
return recvd;
}
Hi,
On Thu, 2022-07-28 at 15:48 +0200, Mark Wielaard wrote:
> With GCC 12.1.1, glibc 2.3a, -fsanitize=undefined and
> -D_FORTIFY_SOURCE=3 we get the following error message:
>
> In file included from /usr/include/ar.h:22,
> from ../libelf/libelfP.h:33,
> from core-file.c:31:
> In function ‘pread’,
> inlined from ‘pread_retry’ at ../lib/system.h:188:21,
> inlined from ‘elf_begin_rand’ at core-file.c:86:16,
> inlined from ‘core_file_read_eagerly’ at core-file.c:205:15:
> /usr/include/bits/unistd.h:74:10: error: ‘__pread_alias’ writing 58
> or more bytes into a region of size 10 overflows the destination [-
> Werror=stringop-overflow=]
> 74 | return __glibc_fortify (pread, __nbytes, sizeof (char),
> | ^~~~~~~~~~~~~~~
> /usr/include/ar.h: In function ‘core_file_read_eagerly’:
> /usr/include/ar.h:41:10: note: destination object ‘ar_size’ of size
> 10
> 41 | char ar_size[10]; /* File size, in ASCII
> decimal. */
> | ^~~~~~~
> /usr/include/bits/unistd.h:50:16: note: in a call to function
> ‘__pread_alias’ declared with attribute ‘access (write_only, 2, 3)’
> 50 | extern ssize_t __REDIRECT (__pread_alias,
> | ^~~~~~~~~~
> cc1: all warnings being treated as errors
>
> The warning disappears when dropping either -fsanitize=undefined
> or when using -D_FORTIFY_SOURCE=2. It looks like a false positive.
> But I haven't figured out how/why it happens.
>
> The code is a little tricky to proof correct though. The ar_size
> field is a not-zero terminated string ASCII decimal, right-paddedr
> with spaces. Which is then converted with strtoll. Relying on the
> fact that the struct ar_hdr is zero initialized, so there will be
> a zero byte after the ar_size field.
>
> Rewrite the code to just use a zero byte terminated char array.
> Which is much easier to reason about. As a bonus the error
> disappears.
The try build turned out green (ppc64le and s390x were red before)
except for the centos7 builder where the native-biarch-core failed
(this is a flaky test apparently because of a kernel issue dumping
biarch cores?) An explicit rebuild made all tests PASS.
So I have pushed this to get all our builders green again.
Cheers,
Mark
@@ -1,3 +1,8 @@
+2022-07-28 Mark Wielaard <mark@klomp.org>
+
+ * core-file.c (elf_begin_rand): Replace struct ar_hdr h with
+ a char ar_size[AR_SIZE_CHARS + 1] array to read size.
+
2022-07-18 Shahab Vahedi <shahab@synopsys.com>
* debuginfod-client.c (dwfl_get_debuginfod_client stub):
@@ -75,26 +75,32 @@ elf_begin_rand (Elf *parent, off_t offset, off_t size, off_t *next)
from the archive header to override SIZE. */
if (parent->kind == ELF_K_AR)
{
- struct ar_hdr h = { .ar_size = "" };
-
- if (unlikely (parent->maximum_size - offset < sizeof h))
+ /* File size, in ASCII decimal, right-padded with ASCII spaces.
+ Max 10 characters. Not zero terminated. So make this ar_size
+ array one larger and explicitly zero terminate it. As needed
+ for strtoll. */
+ #define AR_SIZE_CHARS 10
+ char ar_size[AR_SIZE_CHARS + 1];
+ ar_size[AR_SIZE_CHARS] = '\0';
+
+ if (unlikely (parent->maximum_size - offset < sizeof (struct ar_hdr)))
return fail (ELF_E_RANGE);
if (parent->map_address != NULL)
- memcpy (h.ar_size, parent->map_address + parent->start_offset + offset,
- sizeof h.ar_size);
+ memcpy (ar_size, parent->map_address + parent->start_offset + offset,
+ AR_SIZE_CHARS);
else if (unlikely (pread_retry (parent->fildes,
- h.ar_size, sizeof (h.ar_size),
+ ar_size, AR_SIZE_CHARS,
parent->start_offset + offset
+ offsetof (struct ar_hdr, ar_size))
- != sizeof (h.ar_size)))
+ != AR_SIZE_CHARS))
return fail (ELF_E_READ_ERROR);
- offset += sizeof h;
+ offset += sizeof (struct ar_hdr);
char *endp;
- size = strtoll (h.ar_size, &endp, 10);
- if (unlikely (endp == h.ar_size)
+ size = strtoll (ar_size, &endp, 10);
+ if (unlikely (endp == ar_size)
|| unlikely ((off_t) parent->maximum_size - offset < size))
return fail (ELF_E_INVALID_ARCHIVE);
}